Peak vs. Off-Peak · Flagship-Gewichte · gleiches Basis-Nachtraining · Preismacht
Wer eine DeepSeek-API-Rechnung zahlt, Qwen3.8-Max-Gewichte laden oder GLM-5.3 im Coding-Bench vergleichen will, liest die letzten fünf Tage leicht falsch. Schlagzeilen nennen 1.100 %, volle Open Source und einen Sprung. Dieser Text legt den 12.–17. August als Zeitlinie, Preistabelle und Lizenztext. Fazit: Chinas Labs wechseln vom reinen Preiswettbewerb zur Preismacht. Drei Datentabellen, drei Strategien, sechs Prüfschritte. DSGVO-relevante Hosting-Fragen bleiben bei Selbsthosting auf eigener Hardware klarer als bei undurchsichtigen Weiterverkäufern.
In five days, three of China's top labs made moves that look contradictory. DeepSeek raised API prices by as much as 1,100% on certain tiers. Alibaba open-weighted a 2.4-trillion-parameter flagship it had never released. Zhipu shipped GLM-5.3 and lifted coding benches by roughly 6x on the same base model. Avoid these traps first:
Treat "11x", "1,100%", and "350%" as one number: All three are true. They map to different line items — cache-hit input, output, and cache-miss input. The 1,100% headline is peak-hour cache-hit input only.
Read "open weights" as Apache 2.0: Qwen3.8-2.4T-A95B ships under a custom Qwen3.8-Max License. A Model-as-a-Service or AI Work Assistant business over $50 million in any 12-month window must negotiate a separate commercial license.
Assume GLM-5.3 is a new, larger base: It reuses GLM-5.2's 743B foundation. No retraining. The lift comes from scaling reinforcement-learning environments in post-training.
Keep the "Chinese model = cheapest" default: Off-peak V4-Pro still undercuts Claude Opus 5. International Qwen3.8-Max and GPT-5.6 Luna now undercut DeepSeek off-peak on at least one dimension.
Promote chip claims, a Cursor bug, or export-control rumors to fact: Zhenwu M890, GLM-5.3's reported Cursor finding, and Ministry of Commerce retaliation remain unverified or vendor-sourced in the source notes.
| Date | Event |
|---|---|
| Jul 16, 2026 | Moonshot AI open-weights Kimi K3 (2.8T parameters), drawing US security scrutiny |
| Aug 2–3, 2026 | Alibaba previews, then launches, Qwen3.8-Max as a hosted API |
| Aug 10, 2026 | Meta releases Muse Glimmer (30B, Apache 2.0) and teases open weights for flagship Muse Spark 1.2 |
| Aug 12, 2026 | Alibaba publishes Qwen3.8-2.4T-A95B open weights on Hugging Face / ModelScope; xAI ships Grok 4.6 |
| Aug 13, 2026 | DeepSeek-V4-Pro goes GA and announces a price increase effective Aug 17; Google ships discounted Gemini 3.7 Flash |
| Aug 14, 2026 | Zhipu ships GLM-5.3, reusing GLM-5.2's 743B base |
| Aug 17, 2026, 00:00 Beijing time | DeepSeek's new pricing takes effect |
Zoom out. On Jul 30, OpenAI cut GPT-5.6 Luna by 80%. On Aug 6–7 it made Luna the free default with unlimited text chats. Chinese labs raised prices and opened flagship weights. US labs cut prices and went free at the consumer layer. Same week. Same fight. Two sides.
Do not switch vendors on a headline. Run this check first.
Split the price sheet: Cache-hit input, cache-miss input, and output. Then peak vs off-peak. The ~1,100% figure is V4-Pro peak cache-hit input (¥0.025 → ¥0.30). Output at 350% and cache-miss input at 200% move most real bills.
Map your hours: Peak is 9am–12pm and 2pm–6pm Beijing time. Heavy users with high cache hits and mostly off-peak traffic see closer to a 1.8x bill, per third-party modeling in the source notes. Peak plus low cache hits tracks the headline.
Read the LICENSE file, not the announcement thread: Qwen3.8-Max License has no geographic ban. Triggers: MaaS or AI Work Assistant revenue over $50 million in any consecutive 12 months needs a separate license; 100M+ MAU or $20M+ monthly revenue must display the model name. Claims of a US/EU/UK/Korea download ban are false.
Separate vendor benches from third-party reruns: GLM-5.3's Terminal-Bench 3.0 jump (4.6% → 28.3%) is Zhipu's own number. No independent rerun is published. It still trails GPT-5.6 Sol (34.6%) and Claude Fable 5 (33.7%). Top open-weight tier, not an outright frontier win.
Build a layered price table: V4-Pro peak ~$1.26 in / $3.78 out; off-peak ~$0.63 / $1.89; Qwen3.8-Max international $2 / $6; Luna $0.20 / $1.20; Claude Opus 5 implied ~$5 / $25. RMB-to-USD at about ¥7.15/$1. Official notices win if they disagree.
Shift load and, if needed, host locally: "More flexible workload scheduling" means peak compute is scarce. Batch work into off-peak hours. Teams that self-host open weights or run 24/7 agents can put inference on predictable Apple Silicon. Budget against Mac Mini M4 rental pricing.
Note: Recheck official pricing and license text before you republish. Cache-hit input rose the most (up to about 12x / 1,100%+). Absolute dollars on that line stay small.
Peak hours are 9am–12pm and 2pm–6pm Beijing time. Source: DeepSeek's official announcement, cross-checked against Wall Street CN, IT Home, and V2EX.
| Billing item (per 1M tokens) | Old price | New off-peak | New peak | Peak increase |
|---|---|---|---|---|
| V4-Flash cache hit (input) | ¥0.02 | ¥0.05 | ¥0.10 | ~400% |
| V4-Flash cache miss (input) | ¥1.0 | ¥1.5 | ¥3.0 | 200% |
| V4-Flash output | ¥2.0 | ¥4.5 | ¥9.0 | 350% |
| V4-Pro cache hit (input) | ¥0.025 | ¥0.15 | ¥0.30 | ~1,100% |
| V4-Pro cache miss (input) | ¥3.0 | ¥4.5 | ¥9.0 | 200% |
| V4-Pro output | ¥6.0 | ¥13.5 | ¥27.0 | 350% |
The 1,100% figure applies to peak-hour cache-hit input — the tier that started closest to free. Output, which dominates most real bills, rose 350%. Independent cost modeling cited in the source notes found a realistic heavy-usage workload (roughly 84M tokens/month, mostly off-peak, half cache hits) sees a bill increase closer to 1.8x.
| Spec | Detail |
|---|---|
| Parameters | 2.4T total, 95B active per token (MoE, 512 experts, 10 routed + 1 shared) |
| Context window | 262,144 tokens native (open checkpoint), extendable to ~1.01M; hosted Max defaults to 1M |
| Release cadence | Preview Aug 2 → API live Aug 3 → open weights Aug 12 |
| API pricing (international) | $2/M input, $6/M output |
| License | Not Apache 2.0 — a custom Qwen3.8-Max License |
| Why it matters | First time Alibaba has open-weighted a Max-tier flagship; Qwen3.5 / 3.6 / 3.7 Max stayed API-only |
| Benchmark | GLM-5.2 | GLM-5.3 | Change |
|---|---|---|---|
| Terminal-Bench 3.0 | 4.6% | 28.3% | +23.7 pts |
| DeepSWE v1.1 | 46.2% | 66.9% | +20.7 pts |
| Agents' Last Exam (CLI) | 23.8% | 28.5% | +4.7 pts |
| CyberGym | 77.2% | 84.5% | +7.3 pts |
| AutomationBench | 26.2% | 48.2% | +22.0 pts |
Caution: These are Zhipu's own reported numbers. No independent third-party re-run has been published. GLM-5.3 still trails GPT-5.6 Sol (34.6%) and Claude Fable 5 (33.7%) on Terminal-Bench 3.0.
The easy misread is "China's cheapest model finally caved to margin pressure." The structure says the opposite. DeepSeek is making compute constraints visible on the price sheet. Flat, always-cheap pricing worked as a customer-acquisition tool while GPU capacity kept pace. Usage grew. Capacity did not. "Encouraging more flexible workload scheduling" is corporate-speak for scarce peak-hour compute.
One detail most international coverage missed: at peak hours, DeepSeek's official API now prices above several third-party resellers (GMI Cloud, Novita, and others still list V4 Pro below the new official peak). The assumption that the official API is always the cheapest way to run DeepSeek is broken for the first time.
This is not a gift. Alibaba published the full 2.4T checkpoint and attached a custom license — not the Apache 2.0 used for smaller Qwen models. Any Model-as-a-Service or AI Work Assistant business earning over $50 million in any 12-month period must negotiate a separate commercial license. Products with 100M+ monthly active users or $20M+ in monthly revenue must display the model name.
The bet: give away weights to win developers, especially overseas, where a "made-in-China model" still meets hesitation. Keep leverage over the few firms that can build a competing inference business on top. That is a different bet than Meta's Muse Glimmer, which ships under unrestricted Apache 2.0.
Kill one rumor: claims that the license bans downloads from the US, EU, UK, and South Korea are false. The published text has no geographic clause. In a release cycle this fast, the LICENSE file is the primary source.
The method matters more than the score. Same 743B base as GLM-5.2. No retraining. Terminal-Bench 3.0 moves from 4.6% to 28.3% — about 6x — from scaling RL environments in post-training. Pretraining scaling laws are showing diminishing returns. Post-training RL scale is now an independent lever with a lower cost floor than retraining a new foundation model. Mid-tier labs without OpenAI-scale compute can still close the gap on agentic and coding benches.
Three labs. Three plays. One signal: China's AI labs are shifting from competing on price alone to competing on pricing power.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Open weights? |
|---|---|---|---|
| DeepSeek V4-Pro (peak) | ¥9.0 (~$1.26) | ¥27.0 (~$3.78) | No |
| DeepSeek V4-Pro (off-peak) | ¥4.5 (~$0.63) | ¥13.5 (~$1.89) | No |
| Qwen3.8-Max (international API) | $2.00 | $6.00 | Yes (custom license) |
| OpenAI GPT-5.6 Luna | $0.20 | $1.20 | No |
| Claude Opus 5 (implied, per Alibaba's comparison ratio) | ~$5.00 | ~$25.00 | No |
RMB-to-USD at about ¥7.15/$1, approximate. Short answer: no. Off-peak V4-Pro still sits well below Claude Opus 5. It is no longer the outright cheapest option. International Qwen3.8-Max and Luna both undercut DeepSeek's off-peak rate. "Chinese model = cheapest model" held for most of 2025 and early 2026. It is not a safe assumption anymore.
China's top labs have shipped major releases at a pace domestic financial media now calls "three model updates a week" (一周三更) — DeepSeek, Alibaba, and Zhipu, plus Moonshot's Kimi K3 (open-weighted Jul 16, 2.8T) and MiniMax H3. Chinese coverage frames this as open-weight releases forcing a global repricing of the industry. English-language coverage has mostly treated the same events as isolated product news.
US labs run the opposite play at the consumer layer. OpenAI cut its cheapest tier 80% on Jul 30, then made that model free and unlimited a week later. Google shipped a coding-focused model at half the price of its three-week-old predecessor on Aug 13. Chinese labs open-weight flagships and introduce tiered, higher pricing on the compute-constrained top end. US labs race toward free and cheap at the consumer end. Both are real. They optimize different parts of the funnel.
There is also a geopolitical layer. Moonshot's Kimi K3 already drew US security scrutiny. Some analysts read Alibaba's choice of this window to open-weight a 2.4T flagship as a move to lock in international mindshare and a "technological parity" narrative before any potential regulatory tightening. That is an informed interpretation, not a confirmed fact. It is hard to see if you only read English-language tech press.
"Chinese model = cheapest model" held for most of 2025 and early 2026. After this hike, it is not a safe default.
Public-cloud VMs often add performance tax, weak Apple Silicon / Metal support, and shaky long-run stability for teams that want to self-host a 2.4T-class checkpoint, shift Agent load off peak APIs, or stop treating eval boxes as throwaway sandboxes. For a more stable production path that fits iOS CI/CD and AI Agent automation, VpsMesh Mac Mini cloud rental is usually the better fit: physical Apple Silicon, predictable monthly rent, and a place to pull inference off the peak-hour invoice. See Mac Mini M4 rental pricing, the help center, or order a cloud Mac.
Sources: DeepSeek's official pricing announcement, cross-checked against Wall Street CN, IT Home, AIGC.cn, and V2EX · Alibaba Qwen repositories on Hugging Face / ModelScope and South China Morning Post reporting on license terms · Zhipu (Z.ai) GLM-5.3 technical page, plus VentureBeat and StableLearn · Meta AI Research official blog and VentureBeat on Muse Glimmer · Yicai and Sohu Finance on the pacing of China's open-weight cycle. Figures reflect public information as of publication. Verify the latest official pricing and license terms before republishing. Domestic-chip claims, the Cursor vulnerability report, and export-control rumors remain unconfirmed.
Its off-peak rate is still cheaper than Claude Opus 5, but it is no longer the single cheapest option overall. OpenAI GPT-5.6 Luna ($0.20 / $1.20 per million tokens) and Alibaba's international Qwen3.8-Max pricing ($2 / $6) now undercut DeepSeek's new off-peak rates on at least one dimension. DeepSeek is still relatively cheap for a frontier-class model. It is not the outright cheapest anymore.
Yes for most use cases. Personal projects and internal enterprise use are unaffected. The catch applies only if you run a Model-as-a-Service or AI Work Assistant business that has earned over $50 million in any consecutive 12-month period. That tier requires a separate commercial license from Alibaba.
No. That claim circulated online but is false. The published license contains no geographic restriction of any kind. The restrictions are revenue-based, not tied to where you or your users are located.
Nothing at the base-model level. Both use the same 743-billion-parameter foundation model. The performance gains (roughly 6x on Terminal-Bench 3.0) come entirely from scaling up reinforcement learning during post-training, with no retraining of the base model.
Not yet. Muse Glimmer is a 30B distilled model, not Meta's real flagship. CEO Mark Zuckerberg has said open weights for the larger, closed Muse Spark 1.2 are coming "soon." If that happens, it would be the first US flagship-tier model released openly. As of this writing, treat it as a stated intention. To host open weights or off-peak agents on physical Apple Silicon, see Mac Mini M4 rental pricing and the help center.